A clustering algorithm of cross-emerging-cluster large margin Bayesian fuzzy clustering(CECLM-BFC) is proposed for antecedent parameter learning. The CECLM-BFC algorithm considers repulsed force of clustering centers belonging to heterogeneous samples, and makes the maximum distances between clustering centers. A particle filter method is performed aliematively in different samples to obtain the optimal model parameters, involving the cluster numbers, fuzzy memberships and clustering centers. The learning strategy in consequent parameter learning is based on the maximum classification separation, and a Bayesian Mamdani-Assilan type fuzzy system(BMA-FS) is proposed. The experiment results show the effectiveness of the BMA-FS on classification accuracy and the number of fuzzy rules, and the obtained rules also have high interpretability.